UPPER BOUNDS FOR PARTIAL SPREADS
SASCHA KURZ?
ABSTRACT. Apartialt-spreadinFnq is a collection oft-dimensional subspaces with trivial intersection such that each non-zero vector is covered at most once. We present some improved upper bounds on the maximum sizes.
Keywords:Galois geometry, partial spreads, constant dimension codes, and vector space partitions MSC:51E23; 05B15, 05B40, 11T71, 94B25
1. INTRODUCTION
Let q > 1 be a prime power andn a positive integer. Avector space partitionP of Fnq is a collection of subspaces with the property that every non-zero vector is contained in a unique member ofP. IfPcontainsmd
subspaces of dimensiond, thenP is of typekmk. . .1m1. We may leave out some of the cases withmd = 0.
Subspaces of dimension1are calledholes. If there is at least one non-hole, thenPis called non-trivial.
Apartialt-spreadinFnq is a collection oft-dimensional subspaces such that the non-zero vectors are covered at most once, i.e., a vector space partition of typetmt1m1. ByAq(n,2t;t)we denote the maximum value of mt1. Writingn = kt+r, withk, r ∈ N0andr ≤t−1, we can state that forr ≤ 1orn ≤ 2tthe exact value ofAq(n,2t;t)was known for more than forty years [1]. Via a computer search the casesA2(3k+ 2,6; 3) were settled in 2010 by El-Zanati et al. [5]. In 2015 the caseq=r= 2was resolved by continuing the original approach of Beutelspacher [13], i.e., byconsideringthe set of holes in(n−2)-dimensional subspaces and some averaging arguments. Very recently, N˘astase and Sissokho found a very clear generalized averaging method for the number of holes in(n−j)-dimensional subspaces, wherej ≤ t−2, and generalq, see [14]. Their Theorem 5 determines the exact values of Aq(kt+r,2t;t)in all cases wheret > r
1
q := qq−1r−1. Here, we streamline and generalize their approach leading to improved upper bounds onAq(n,2t;t), c.f. [15].
2. SUBSPACES WITH THE MINIMUM NUMBER OF HOLES
Definition 2.1. A vector space partitionP ofFnq hashole-type(t, s, m1), if it is of typetmt. . . sms1m1, for some integersn > t≥s≥2,mi∈N0fori∈ {1, s, . . . , t}, andPis non-trivial.
Lemma 2.2. (C.f.[14, Proof of Lemma 9].) LetP be a non-trivial vector space partition ofFnq of hole-type (t, s, m1)andl, x∈N0withPt
i=smi=lqs+x.PH={U∩H : U ∈ P}is a vector space partition of type tm0t. . .(s−1)m0s−11m01, for a hyperplaneH withmb1holes (ofP). We havemb1 ≡ m1+x−1q (modqs−1). If s >2, thenPHis non-trivial andm01=mb1.
PROOF. IfU ∈ P, thendim(U)−dim(U∩H)∈ {0,1}for an arbitrary hyperplaneH. SinceP is non-trivial, we haven≥s. Fors >2, counting the1-dimensional subspaces ofFnq andH, viaPandPH, yields
(lqs+x)· s
1
q
+aqs+m1= n
1
q
and (lqs+x)· s−1
1
q
+a0qs−1+mb1= n−1
1
q
for somea, a0∈N0. Since1 +q·n−1 1
q−n 1
q = 0we conclude1 +qmb1−m1−x≡0 (modqs). Thus, Z3mb1≡m1+x−1q (mod qs−1). Fors= 2we have
lq2+x
·(q+ 1) +aq2+m1= n
1
q
and lq2+x
·1 +a0q+mb1= n−1
1
q
?The work of the author was supported by the ICT COST Action IC1104 and grant KU 2430/3-1 – Integer Linear Programming Models for Subspace Codes and Finite Geometry from the German Research Foundation.
1The more general notationAq(n,2t−2w;t)denotes the maximum cardinality of a collection oft-dimensional subspaces, whose pair- wise intersections have a dimension of at mostw. Those objects are calledconstant dimension codes, see e.g. [6]. For known bounds, we re- fer tohttp://subspacecodes.uni-bayreuth.de[10] containing also the generalization tosubspace codesof mixed dimension.
1
2 SASCHA KURZ
leading to the same conclusionmb1≡ m1+x−1q (modqs−1).
Lemma 2.3. (C.f.[14, Proof of Lemma 9].) LetP be a vector space partition ofFnq of hole-type(t, s, m1), l, x ∈ N0 withPt
i=smi = lqs+x, andb, c ∈ Zwithm1 = bqs+c ≥ 1. Ifx ≥ 1, then there exists a hyperplaneHb withmb1=bbqs−1+bcholes, wherebc:= c+x−1q ∈Zandb >bb∈Z.
PROOF. Apply Lemma 2.2 and observem1≡c (modqs). Let the number of holes inHb be minimal. Then,
mb1≤average number of holes per hyperplane=m1· n−1
1
q
/ n
1
q
< m1
q . (1)
Assumingbb≥byieldsqmb1≥q·(bqs−1+bc) =bqs+c+x−1≥m1, which contradicts Inequality (1).
Corollary 2.4. Using the notation from Lemma 2.3, letP be a non-trivial vector space partition withx≥1 andf be the largest integer such thatqf dividesc. For each0≤j≤s−max{1, f}there exists an(n−j)- dimensional subspaceU containingmb1holes withmb1≡bc (modqs−j)andmb1≤(b−j)·qs−j+bc, where bc=c+[j1]q·(x−1)
qj .
Proof. Observemb1≡c6≡0 (modqs−j), i.e.,mb1≥1, for allj < s−f. Lemma 2.5. LetP be a non-trivial vector space partition of typetmt1m1 ofFnq withmt = lqt+x, where l = qn−tqt−1−qr, x ≥ 2, t = r
1
q + 1−z+u > r, qf|x−1, qf+1 - x−1, and f, u, z, r, x ∈ N0. For max{1, f} ≤y≤tthere exists a(n−t+y)-dimensional subspaceU withL≤(z+y−1)qy+wholes, where w=−(x−1)y
1
qandL≡w(mod qy).
PROOF. Apply Corollary 2.4 withs=t,j =t−y,b=r 1
q, andm1=r 1
qqt−t 1
q(x−1).
Lemma 2.6. LetPbe a vector space partition ofFnq withc≥1holes andaidenote the number of hyperplanes containingiholes. Then,Pc
i=0ai =n 1
q,Pc
i=0iai=c·n−1
1
qandPc
i=0i(i−1)ai=c(c−1)·n−2
1
q. PROOF. Double-count the incidences of the tuples(H),(B1, H), and(B1, B2, H), whereH is a hyperplane
andB16=B2are points contained inH.
Lemma 2.7. Let ∆ = qs−1,m ∈ Z, andP be a vector space partition ofFnq of hole-type (t, s, c). Then, τq(c,∆, m)·q∆n−22 −m(m−1)≥0, where
τq(c,∆, m) =m(m−1)∆2q2−c(2m−1)(q−1)∆q+c(q−1)
c(q−1) + 1 . PROOF. Consider the three equations from Lemma 2.6. (c−m∆)
c−(m−1)∆
times the first minus
2c−(2m−1)∆−1
times the second plus the third equation, and then divided by∆2/(q−1), gives
(q−1)·
bc/∆c
X
h=0
(m−h)(m−h−1)ac−h∆=τq(c,∆, m)·qn−2
∆2 −m(m−1)
due to Lemma 2.2. Finally, we observeai≥0and(m−h)(m−h−1)≥0for allm, h∈Z. Lemma 2.8. For integersn > t≥s≥2and1≤i≤s−1, there exists no vector space partitionPofFnq of hole-type(t, s, c), wherec=i·qs−s
1
q+s−1.2
PROOF. Assume the contrary and apply Lemma 2.7 withm=i(q−1). Settingy =s−1−iand∆ =qs−1 we compute
τq(c,∆, m) = −q∆(y(q−1) + 2) + (s−1)2q2−q(s−1)(2s−5) + (s−2)(s−3).
Usingy≥0we obtainτ2(c,∆, m)≤s2+s−2s+1<0. Fors= 2, we haveτq(c,∆, m) =−q2+q <0and forq, s >2we haveτq(c,∆, m)≤ −2qs+ (s−1)2q2<0. Thus, Lemma 2.7 yields a contradiction.
2For more general non-existence results of vector space partitions see e.g. [9, Theorem 1] and the related literature.
UPPER BOUNDS FOR PARTIAL SPREADS 3
Theorem 2.9. (C.f. [14, Lemma 10].) For integersr ≥ 1,k ≥ 2, u ≥ 0, and0 ≤ z ≤ r 1
q/2 witht = r
1
q+ 1−z+u > rwe haveAq(n,2t;t)≤lqt+ 1 +z(q−1), wherel= qn−tqt−1−qr andn=kt+r.
PROOF. Apply Lemma 2.5 withx= 2 +z(q−1)andy = z+ 1. Ifz = 0, thenL < 0. Forz ≥ 1, apply
Lemma 2.8. Thus,Aq(n,2t;t)≤lqt+x−1.
The known constructions for partialt-spreads giveAq(kt+r,2t;t)≥lqt+ 1, see e.g. [1] (or [13] for an interpretation using the more general multilevel construction for subspace codes). Thus, Theorem 2.9 is tight fort≥r
1
q+ 1, c.f. [14, Theorem 5].
Theorem 2.10. (C.f.[15, Theorem 6,7].) For integersr≥1,k≥2,y≥max{r,2},z≥0withλ=qy,y≤t, t=r
1
q+ 1−z > r,n=kt+r, andl= qn−tqt−1−qr, we have Aq(n,2t;t)≤lqt+
λ−1
2−1 2
p1 + 4λ(λ−(z+y−1)(q−1)−1)
. PROOF. From Lemma 2.5 we concludeL≤(z+y−1)qy−(x−1)y
1
qandL≡ −(x−1)y 1
q (modqy)for the number of holes of a certain(n−t+y)-dimensional subspaceUofFnq.PU :={P∩U |P∈ P}is of hole- type(t, y, L)ify ≥2. Next, we will show thatτq(c,∆, m)≤0, where∆ =qy−1andc =iqy−(x−1)y
1
q
with1≤i≤z+y−1, for suitable integersxandm. Note that, in order to apply Lemma 2.5, we have to satisfy x≥2andy≥ffor all integersfwithqf|x−1. Applying Lemma 2.7 then gives the desired contradiction, so thatAq(n,2t;t)≤lqt+x−1.
We choose3m=i(q−1)−(x−1) + 1, so thatτq(c,∆, m) =x2−(2λ+ 1)x+λ(i(q−1) + 2). Solving τq(c,∆, m) = 0forxgivesx0 = λ+ 12 ±12θ(i), whereθ(i) = p
1−4iλ(q−1) + 4λ(λ−1). We have τq(c,∆, m)≤0for|2x−2λ−1| ≤θ(i). We need to find an integerx≥2such that this inequality is satisfied for all1≤i≤z+y−1. The strongest restriction is attained fori=z+y−1. Sincez+y−1≤r
1
q and u=qy ≥qr, we haveθ(i)≥θ(z+y−1)≥1, so thatτq(c,∆, m)≤0forx=
u+12−12θ(z+y−1) . (Observex≤λ+12+12θ(z+y−1)due toθ(z+y−1)≥1.) Sincex≤λ+ 1, we havex−1≤λ=qy, so that qf|x−1impliesf ≤yprovidedx≥2. The latter is true due toθ(z+y−1)≤p
1−4λ(q−1) + 4λ(λ−1)≤ p1 + 4λ(λ−2)<2(λ−1), which impliesx≥3
2
= 2.
So far we have constructed a suitablem∈Zsuch thatτq(c,∆, m)≤0forx=
λ+12−12θ(z+y−1) . Ifτq(c,∆, m)<0, then Lemma 2.7 gives a contradiction, so that we assumeτq(c,∆, m) = 0in the following.
Ifi < z+y−1we haveτq(c,∆, m) <0 due toθ(i) > θ(z+y−1), so that we assumei =z+y−1.
Thus,θ(z+y−1)∈ N0. However, we can writeθ(z+y−1)2 = 1 + 4λ(λ−(z+y−1)(q−1)−1) = (2w−1)2= 1+4w(w−1)for some integerw. Ifw /∈ {0,1}, thengcd(w, w−1) = 1, so that eitherλ=qy|w orλ =qy | w−1. Thus, in any case, w≥ qy, which is impossible since(z+y−1)(q−1) ≥1. Finally, w∈ {0,1}impliesw(w−1) = 0, so thatλ−(z+y−1)(q−1)−1 = 0. Thus,z+y−1 =y
1
q ≥r 1
q
sincey≥r. The assumptionsy≤tandt=r 1
q+ 1−zimplyz+y−1 =r 1
qandy=r. This givest=r,
which is excluded.
Setting y = t in Theorem 2.10 yields [4, Corollary 8], which is based on [3, Theorem 1B]. And indeed, our analysis is very similar to the technique4used in [3]. Compared to [3, 4], the new ingredients essentially are lemmas 2.2 and 2.3, see also [14, Proof of Lemma 9]. [4, Corollary 8], e.g., givesA2(15,12; 6) ≤516, A2(17,14; 7) ≤ 1028, and A9(18,16; 8) ≤ 3486784442, while Theorem 2.10 gives A2(15,12; 6) ≤ 515, A2(17,14; 7)≤1026, andA9(18,16; 8)≤3486784420. Postponing the details and proofs to a more extensive and technical paper [12], we state:
• 24l+ 1≤A2(4k+ 3,8; 4)≤24l+ 4, wherel=24k−124−1−23 andk≥2, e.g.,A2(11,8; 4)≤132;
• 26l+ 1≤A2(6k+ 4,12; 6)≤26l+ 8, wherel= 26k−226−1−24 andk≥2, e.g.,A2(16,12; 6)≤1032;
• 26l+ 1≤A2(6k+ 5,12; 6)≤26l+ 18, wherel=26k−126−1−25 andk≥2, e.g.,A2(17,12; 6)≤2066;
• 34l+ 1≤A3(4k+ 3,8; 4)≤34l+ 14, wherel= 34k−134−1−33 andk≥2, e.g.,A3(11,8; 4)≤2201;
3Solving∂τq(c,∆,m)∂m = 0, i.e., minimizingτq(c,∆, m), yieldsm=i(q−1)−(x−1) +12 +x−1qy . Fory≥rwe can assume x−1< qydue the known constructions for partial spreads, so that up-rounding yields the optimum integer choice. Fory < rthe interval u+12−12θ(i), u+12+12θ(i)
may contain no integer.
4Actually, their analysis grounds on [16] and is strongly related to the classical second-order Bonferroni Inequality [2, 7, 8] in Probability Theory, see e.g. [11, Section 2.5] for another application for bounds on subspace codes.
4 SASCHA KURZ
• 35l+ 1≤A3(5k+ 3,10; 5)≤35l+ 13, wherel=35k−233−1−35 andk≥2, e.g.,A3(13,10; 5)≤6574;
• 35l+ 1≤A3(5k+ 4,10; 5)≤35l+ 44, wherel=35k−135−1−34 andk≥2, e.g.,A3(14,10; 5)≤19727;
• 36l+ 1≤A3(6k+ 4,12; 6)≤36l+ 41, wherel=36k−236−1−34 andk≥2, e.g.,A3(16,12; 6)≤59090;
• 36l+1≤A3(6k+5,12; 6)≤36l+133, wherel= 36k−136−1−35 andk≥2, e.g.,A3(17,12; 6)≤177280;
• 37l+ 1≤A3(7k+ 4,14; 7)≤37l+ 40, wherel= 37k−337−1−34 andk≥2, e.g.,A3(18,14; 7)≤177187;
• 45l+ 1≤A4(5k+ 3,10; 5)≤45l+ 32, wherel=45k−245−1−43 andk≥2, e.g.,A4(13,10; 5)≤65568;
• 46l+ 1≤A4(6k+ 3,12; 6)≤46l+ 30, wherel= 46k−346−1−43 andk≥2, e.g.,A4(15,12; 6)≤262174;
• 46l+1≤A4(6k+5,12; 6)≤46l+548, wherel=46k−146−1−45 andk≥2, e.g.,A4(17,12; 6)≤4194852;
• 47l+1≤A4(7k+4,14; 7)≤47l+128, wherel=47k−347−1−44 andk≥2, e.g.,A4(18,14; 7)≤4194432;
• 55l+ 1≤A5(5k+ 2,10; 5)≤55l+ 7, wherel= 55k−355−1−52 andk≥2, e.g.,A5(12,10; 5)≤78132;
• 55l+1≤A5(5k+4,10; 5)≤55l+329, wherel=55k−155−1−54 andk≥2, e.g.,A5(14,10; 5)≤1953454;
• 75l+ 1 ≤ A7(5k+ 4,10; 5) ≤ 75l+ 1246, wherel = 75k−175−1−72 andk ≥ 2, e.g., A7(14,10; 5) ≤ 40354853;
• 84l+ 1≤A8(4k+ 3,8; 4)≤84l+ 264, wherel= 84k−184−1−83 andk≥2, e.g.,A8(11,8; 4)≤2097416;
• 85l+1≤A8(5k+2,10; 5)≤85l+25, wherel=85k−385−1−82 andk≥2, e.g.,A8(12,10; 5)≤2097177;
• 86l+1≤A8(6k+2,12; 6)≤86l+21, wherel= 86k−486−1−82andk≥2, e.g.,A8(14,12; 6)≤16777237;
• 93l+ 1≤A9(3k+ 2,6; 3)≤93l+ 41, wherel= 93k−193−1−92 andk≥2, e.g.,A9(8,6; 3)≤59090;
• 95l+ 1 ≤ A9(5k+ 3,10; 5) ≤ 95l+ 365, wherel = 95k−295−1−93 andk ≥ 2, e.g., A9(13,10; 5) ≤ 43047086;
c.f. the web-page mentioned in footnote 1 for more numerical values and comparisons of the different upper bounds.
REFERENCES
[1] A. Beutelspacher,Partial spreads in finite projective spaces and partial designs, Mathematische Zeitschrift145(1975), no. 3, 211–229.
[2] C.E. Bonferroni,Teoria statistica delle classi e calcolo delle probabilit`a, Libreria internazionale Seeber, 1936.
[3] R.C. Bose and K.A. Bush,Orthogonal arrays of strength two and three, The Annals of Mathematical Statistics23(1952), 508–524.
[4] D.A. Drake and J.W. Freeman,Partialt-spreads and group constructible(s, r, µ)-nets, Journal of Geometry13(1979), no. 2, 210–216.
[5] S. El-Zanati, H. Jordon, G. Seelinger, P. Sissokho, and L. Spence,The maximum size of a partial3-spread in a finite vector space over GF(2), Designs, Codes and Cryptography54(2010), no. 2, 101–107.
[6] T. Etzion and L. Storme,Galois geometries and coding theory, Designs, Codes and Cryptography78(2016), no. 1, 311–350.
[7] J. Galambos,Bonferroni inequalities, The Annals of Probability5(1977), no. 4, 577–581.
[8] J. Galambos and I. Simonelli,Bonferroni-type inequalities with applications, Springer Verlag, 1996.
[9] O. Heden,On the length of the tail of a vector space partition, Discrete Mathematics309(2009), no. 21, 6169–6180.
[10] D. Heinlein, M. Kiermaier, S. Kurz, and A. Wassermann,Tables of subspace codes, University of Bayreuth, 2015, available at http://subspacecodes.uni-bayreuth.de and http://arxiv.org/abs/1601.02864.
[11] T. Honold, M. Kiermaier, and S. Kurz,Constructions and bounds for mixed-dimension subspace codes, Advances in Mathematics of Communication10(2016), no. 3, 649–682.
[12] T. Honold, M. Kiermaier, and S. Kurz,Partial spreads and vector space partitions, arXiv preprint 1611.06328 (2016).
[13] S. Kurz,Improved upper bounds for partial spreads, to appear in Designs, Codes and Cryptography, doi: 10.1007/s10623-016-0290-8, arXiv preprint 1512.04297 (2015).
[14] E. N˘astase and P. Sissokho,The maximum size of a partial spread in a finite projective space, arXiv preprint 1605.04824 (2016).
[15] E. N˘astase and P. Sissokho,The maximum size of a partial spread II: Upper bounds, arXiv preprint 1606.09208 (2016).
[16] R.L. Plackett and J.P. Burman,The design of optimum multifactorial experiments, Biometrika33(1946), no. 4, 305–325.
DEPARTMENT OFMATHEMATICS, UNIVERSITY OFBAYREUTH, 95440 BAYREUTH, GERMANY E-mail address:sascha.kurz@uni-bayreuth.de